betaFeedback

WorksComputer Science

RACE: Ranking-Uncertainty Aware Capacity-Adaptive Customization for Efficient Sequential Recommendation

Abstract

Efficient sequential recommendation has emerged as a promising paradigm for providing real-time personalized service. To deploy models under resource-constrained environment, model compression becomes a necessity, yet existing methods typically apply uniform compression ratios, overlooking the heterogeneity in user intent complexity. This "one-size-fits-all" approach leads to a fundamental Capacity-Complexity Mismatch, where computational resources are either wasted on simple instances or insufficient for complex ones. To address this, we propose RACE (Ranking-uncertainty Aware Capacity-Adaptive Customization for Efficient Sequential Recommendation), a framework that dynamically adapts both model structures and input sequences with uncertainty-aware compression ratios. To ensure capacity-adaptive inference, RACE captures the model's intrinsic uncertainty of its predictions and transforms it into a proactive resource allocator. Specifically, we derive a list-wise ranking-uncertainty metric from the distributional stability of top-KK predictions, which serves as a measurable proxy for instance complexity. A lightweight predictor is then trained via self-supervision to anticipate this uncertainty from early-layer features. Guided by this signal, RACE performs instance-specific slimming in two dimensions: a hypernetwork-style channel sensitivity extractor identifies parameter redundancy, while importance-aware progressive sequence filtering mitigates data redundancy. Extensive experiments on four popular datasets demonstrate that RACE outperforms state-of-the-art baselines on accuracy, efficiency, and adaptability.

CCS Concepts: • Information systems → Recommender systems; Personalization.

Keywords

Sequential Recommendation, Efficient Recommendation, Model Customization

ACM Reference Format:

Tianqi Liu, Kaipu Sun, Kairui Fu, Wenkai Wang, Sihao Liu, Hui Zang, Ran Chen, Tao Wang, and Shengyu Zhang. 2027. RACE: Ranking-Uncertainty Aware Capacity-Adaptive Customization for Efficient Sequential Recommendation. In Proceedings of the Twentieth ACM International Conference on Web Search and Data Mining (WSDM ’27), February 15–19, 2027, Hong Kong, Hong Kong. ACM, New York, NY, USA, 12 pages. https://doi.org/10.1145/3862322.3863404

Introduction

In recent years, sequential recommendation has witnessed remarkable progress driven by the advancement of deep learning [4, 8, 29, 55, 59]. Modern approaches [10, 12, 23, 38, 40, 56, 61] leverage increasingly sophisticated Transformer-based architectures to capture user preference patterns from interaction histories. The prevailing methods deploy models on cloud servers, where centralized computational resources support serving billions of users. Although effective, this approach requires uploading user interaction data to remote servers [60], which raises both network latency issues and privacy concerns [15, 54]. Moreover, maintaining personalized models for individual users is computationally prohibitive at scale [25].

To address these challenges, efficient sequential recommendation has gained increasing attention, leveraging advances in mobile computing to support real-time reranking [14, 15, 45, 46]. While this paradigm mitigates latency and privacy concerns [25, 54], achieving personalization on resource-constrained devices remains challenging. Some prior works adapt models to individual users through on-device finetuning [35, 44, 49]. However, as depicted in Figure 1(a),

Image not extracted

Figure 1.

This work is licensed under a Creative Commons Attribution 4.0 International License. WSDM ’27, Hong Kong, Hong Kong. © 2027 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-XXXX-X/2027/02. https://doi.org/10.1145/3862322.3863404 the required backpropagation imposes significant computational burdens. Additionally, in sustainable deployment scenarios [54], the time-consuming update process forces systems to operate with stale models whenever user interests shift [5, 9, 43], degrading recommendation quality. Other works pursue resource-aware model customization [25, 32]. These methods mainly rely on compression techniques: network pruning removes redundancy by identifying sparse masks [11, 28, 39] (Figure 1(b)), while weight decomposition [47, 48] approximates dense matrices via low-rank factorization. Despite their effectiveness, these compression-based techniques apply static model and input size constraints, overlooking the heterogeneity of user intent complexity [45, 54].

Figure 1 (PDF p. 2). Comparison of on-device recommendation paradigms. (a) On-device Finetuning incurs computational overhead due to backpropagation, and often falls short for capturing shifting interests. (b) On-device Pruning applies a unified-sparsity mask and relies on additional extractors, unable to adapt to the complexity of user intents. (c) Our proposed RACE framework introduces a Ranking-Uncertainty Driven mechanism to dynamically adapt model size to instance complexity. In comparison, RACE achieves sequence and structure slimming and customization in an intrinsic feature reuse and retraining-free manner.

In reality, the complexity of user preferences varies significantly across instances (i.e., the most recent interaction sequences) [37]. For users with stable and explicit interests, a relatively lightweight model may suffice to capture the pattern, while those with volatile or complex interaction histories demand greater model expressiveness. This suggests that a static "one-size-fits-all" architecture can lead to a Capacity-Complexity Mismatch, where resources are either wasted on simple instances or insufficient for hard ones. Achieving capacity-adaptive inference thus becomes essential, yet two critical challenges remain:

Challenge (i): How to decide the ideal capacity allocation strategy? Challenge (ii): How to achieve personalized slimming under the allocated capacity?

The first challenge demands a resource allocation metric that reflects instance complexity. A naive solution might rely on intuitive heuristics such as user history popularity or user interest stability. Unfortunately, as Figure 2 shows, neither metric correlates clearly with model performance, suggesting that instance complexity is not a static input feature, but a dynamic property reflecting the confidence of the model itself. Thus, accurately capturing it requires measuring the model’s internal uncertainty. However, existing

Figure 2. Correlation analysis between model performance and two input-side heuristics: (a) history popularity; (b) interest stability. Neither reliably indicates instance complexity.

model uncertainty methods [20, 50] operate primarily at the point-wise level, quantifying prediction variance for individual items while ignoring interdependencies among candidates. Since ranking tasks fundamentally concern relative ordering rather than absolute scores, capturing uncertainty from a list-wise perspective becomes essential. Moreover, to guide capacity allocation before full inference, such uncertainty should be anticipated from early-layer features rather than computed post-hoc.

The second challenge lies in determining what to slim under the allocated capacity. Due to interest heterogeneity [34, 36, 41, 51], parameter and data sensitivity vary across instances, motivating us to jointly optimize both parameter space (model structure) and data space (input sequences) in an interest-aware manner. To identify redundancy, existing methods often incur additional feature-extraction modules [11], or extra neural-networks for prediction [11, 17, 52]. Although effective, they come with an additional computational burden. We instead seek to reuse intrinsic features and utilize lightweight analysis modules to achieve accuracy improvement and inference speedup with minimal additional cost.

In light of this, we propose RACE, a Ranking-uncertainty Aware Capacity-Adaptive Customization framework for Efficient sequential recommendation. As illustrated in Figure 1(c), RACE establishes a capacity-adaptive customization and slimming for both sequences and model structures without on-device backpropagation. To estimate the model’s inner confidence, we design a list-wise ranking-uncertainty metric that quantifies the structural volatility of predicted lists by jointly modeling the distributional characteristics (means and variances) of top-kk predictions via a negative log-likelihood formulation. We further demonstrate that this metric exhibits a clear negative correlation with recommendation performance, validating it as a principled proxy for instance complexity. Crucially, RACE transforms this uncertainty from a post-hoc performance indicator into a proactive resource allocator through a self-supervised learning mechanism. A lightweight predictor then learns to anticipate ranking-uncertainty from early-layer features, enabling the model to autonomously calibrate its computational budget to instance complexity (Challenge i).

Driven by this uncertainty-aware threshold, RACE executes multi-granularity customization across both parameter and data spaces. Specifically, in the parameter space, we identify interest-compatible sub-networks adhering to an intrinsic feature reuse manner. RACE repurposes the backbone’s hidden states as a dynamic user embedding, which a hypernetwork-style sensitivity extractor then consumes to assess channel sensitivity, enabling personalized channel deactivation at negligible cost. Simultaneously, in the data space, the uncertainty-aware signal governs progressive sequence filtering, which calculates attention-aware relative importance and adaptively retains semantically core interactions while suppressing noisy ones at each block (Challenge ii).

We conduct extensive experiments on four real-world datasets and two representative backbones, SASRec [23] and BSARec [38]. Our main contributions are summarized as follows:

  • We reveal the capacity-complexity mismatch in on-device recommendation and propose RACE, a framework that utilizes uncertainty to adapt computational resources to instance complexity.

  • We establish a ranking-uncertainty driven resource allocation paradigm that transforms list-wise ranking stability into a proactive signal for adaptive inference via self-supervised anticipation.

  • We design multi-granularity customization mechanisms, including hypernetwork-style sub-network identification and progressive sequence filtering, achieving joint optimization of model and data redundancy in a single forward pass.

  • Extensive experiments demonstrate that RACE consistently outperforms state-of-the-art baselines in both accuracy and efficiency while exhibiting strong adaptability.

Related works

Efficient Sequential Recommendation

Deploying models under resource-constrained conditions necessitates model compression [15, 25, 54] on the model side or the input side. On the model side, following the lottery ticket hypothesis, SNIP [28] and GEM [39] identify sparse subnetworks before or at the beginning of training, while PEMN [1] learns binary masks over frozen random weights. STTD [47] and APG [48] instead apply tensor-train decomposition to embeddings and low-rank approximation to dense weights, respectively. Recent advances have shifted toward dynamic, instance-aware model adaptation. DIET [11] and Gater [6] perform importance analysis to identify customized parameters. CHORD [32] uses device-cloud collaboration to generate personalized channel-wise mixed-precision quantization strategies for on-device recommendation, among other device-cloud approaches [33, 57]. On the input side, ATP-LLaVA dynamically removes redundant tokens [53]. PerSRec [58] and CAUSE [31] compress user histories into learnable or category-based tokens, respectively, but retain fixed compression budgets. RACE instead jointly adapts model and input size to instance complexity for more flexible resource allocation.

Model Uncertainty in Recommendation

Uncertainty quantification characterizes predictive reliability in recommender systems [7]. Common point-wise estimators, namely Bayesian neural networks [3], MC Dropout [13], and deep ensembles [27], build predictive distributions from posterior weights, stochastic forward passes, and independently trained models, respectively, summarized by item-level variance or entropy. Pairwise-margin heuristics use score gaps to approximate local relative confidence [42]. In recommendation, these signals mainly serve calibration and uncertainty-aware re-ranking [20, 50], and LiDu [30] targets label-free performance estimation. These works assess or refine predictions, whereas RACE anticipates list-wise uncertainty to guide capacity-adaptive inference.

Method

In this section, we first formalize our problem in § 3.1, then introduce an uncertainty-guided resource allocation mechanism in § 3.2, followed by multi-granularity customization strategies for model structure (§ 3.4) and input sequences (§ 3.5).

Preliminary

In efficient sequential recommendation scenario, we consider a system with an item set I={i1,i2,…,i∣I∣}\mathcal{I} = \{i_1, i_2, \ldots, i_{|\mathcal{I}|}\} and a user set U={u1,u2,…,u∣U∣}\mathcal{U} = \{u_1, u_2, \ldots, u_{|\mathcal{U}|}\}. For each user u∈Uu \in\mathcal{U}, their interaction history is represented as a chronologically ordered sequence Su=[s1u,s2u,…,sTu]S^u = [s_1^u, s_2^u, \ldots, s_T^u], where stu∈Is_t^u \in\mathcal{I} represents an item interacted at timestamp tt, and TT denotes the sequence length.

Our task is to predict the user’s next interaction item sT+1u∈Is_{T+1}^u \in\mathcal{I} from the historical sequence SuS^u with the maximum LL. Given the resource-constrained nature of mobile devices, providing computationally efficient personalized recommendation becomes our critical concern. Our primary objective can be formalized as:

max⁡πu∑u∈UR(πu,M)(Su),\max_{\pi^u} \sum_{u \in\mathcal{U}} R(\pi^u, M)(S^u),
s.t.C≤Cmax⁡,(1)\mathrm{s.t.}\quad C \leq C_{\max}, \tag*{(1)}

where πu={τu,τ′u,mu,κu}\pi^u = \{\tau^u, \tau'^u, \mathbf{m}^u, \kappa^u\} is the instance-specific customization policy that determines: threshold τu,τ′u\tau^u, \tau'^u for channel and sequence pruning intensity, channel masks mu\mathbf{m}^u for structure customization, and sequence filterings κu\kappa^u for sequence customization. The function RR applies πu\pi^u to base model MM for personalized predictions. The constraint C≤Cmax⁡C \leq C_{\max} enforces the computational budget.

Ranking-Uncertainty Guided Resource Allocation

The foundation of adaptive inference lies in measuring model uncertainty to serve as the proxy for instance complexity. We propose a list-wise ranking-uncertainty metric that captures the model’s confidence in its predictions, which then guides our resource allocation strategy through a self-supervised learning mechanism.

Figure 3. Overview of the RACE framework. Our approach consists of three key components: (a) Ranking-Uncertainty Guided Resource Allocation: RACE designs a list-wise ranking-uncertainty metric and transforms it into a proactive instance-complexity proxy to guide resource allocation, utilizing a self-supervised threshold predictor to anticipate ratios early in inference. (b) Interest-Compatible Subnets Identification: Using the predicted ratio, a lightweight hypernetwork-style sensitivity extractor generates channel-wise masks to customize each feed-forward layer. (c) Progressive Sequence Filtering: Using the predicted ratio, RACE calculates relative importance scores to retain semantically relevant items while suppressing noise block by block.

List-Wise Ranking Uncertainty

Traditionally, model uncertainty is measured via point-wise estimation methods, such as Bayesian Neural Networks [3], MC Dropout [13], and Deep Ensembles [27], which calculate the variance of isolated prediction scores. However, dropout methods require multiple forward passes and ensemble approaches necessitate maintaining multiple models, both incurring inference latency and storage burdens prohibitive for real-time service. We therefore adopt a variational Bayesian layer [19] as the prediction head, which maintains weights as W∼N(μW,ΣW)W \sim\mathcal{N}(\mu_W, \Sigma_W) and yields both a predictive mean μ∈R∣I∣\boldsymbol{\mu} \in\mathbb{R}^{|I|} and prediction variance σ2∈R∣I∣\boldsymbol{\sigma}^{2} \in\mathbb{R}^{|I|} for all items in a single forward pass.

Nevertheless, point-wise estimation alone is insufficient, since sequential recommendation is fundamentally a ranking problem where the relative order of items matters more than absolute scores. We therefore lift uncertainty from individual scores to the stability of the predicted Top-KK ordering: rather than measuring how uncertain any single score is, we measure how stable the entire ordering remains under the predictive distribution. Concretely, we treat each pairwise comparison within the Top-KK as a probabilistic event and define the overall ranking uncertainty as the aggregated negative log-likelihood of these events, so that ambiguous orderings contribute more uncertainty than well-separated ones.

This construction reduces to deriving the pairwise outranking probability pi>jp_{i>j} under the model’s own predictive distribution, which the variational Bayesian layer makes available in closed form. We adopt a diagonal parameterization of ΣW\Sigma_W across output dimensions [19], so that the predictive components for different items are conditionally independent given the input. The linear projection z=Wxz = W x then produces jointly Gaussian logits with class-wise independent components: zi∼N(μi,σi2)z_i \sim\mathcal{N}(\mu_i, \sigma_i^2), and the pairwise difference zi−zj∼N(μi−μj,σi2+σj2)z_i-z_j \sim\mathcal{N}(\mu_i-\mu_j, \sigma_i^2+\sigma_j^2). Applying Thurstone’s law of comparative judgment [42] to this Gaussian difference, the outranking probability admits the closed-form expression:

pi>j=P(zi>zj)=Φ(μi−μjσi2+σj2),(2)p_{i>j}=P(z_i>z_j)=\Phi\left(\frac{\mu_i-\mu_j}{\sqrt{\sigma_i^2+\sigma_j^2}}\right), \tag*{(2)}

where Φ(⋅)\Phi(\cdot) is the standard normal CDF. Eq. (2) is thus a direct consequence of the model’s own predictive distribution.

Let K={k1,…,kK}K=\{k_1,\ldots,k_K\} denote the indices of the top-KK items according to zz. The ranking uncertainty is then defined as the negative log-likelihood of all pairwise comparisons within the top-KK:

Urank=−∑i=1K−1∑j=i+1Klog⁡pki>kj.(3)\mathcal{U}_{\mathrm{rank}}=-\sum_{i=1}^{K-1}\sum_{j=i+1}^{K}\log p_{k_i>k_j}. \tag*{(3)}

We set K=20K=20 in our experiments, intentionally larger than the recommendation evaluation cut-off so that Urank\mathcal{U}_{\mathrm{rank}} captures the broader structural stability of the predicted ranking rather than only the very top items. Intuitively, Urank\mathcal{U}_{\mathrm{rank}} approaches zero when the model is confident in its ordering, namely when large score gaps and small variances drive each pi>jp_{i>j} toward 11, and inflates whenever score gaps narrow or variances widen, signaling an ambiguous Top-KK ordering and a more complex instance.

Theoretical Motivation

We now provide the theoretical motivation underlying Urank\mathcal{U}_{\mathrm{rank}}. Two questions need to be addressed: first, why model uncertainty can serve as a signal of instance complexity; second, whether our list-wise construction preserves the probabilistic semantics that makes this signal meaningful.

Calibration makes uncertainty a principled complexity signal. The use of predictive uncertainty as a signal of instance complexity is well-established in machine learning, with roots in curriculum learning [2] and uncertainty quantification [21]. This usage rests on calibration [18, 22]: when a model’s confidence aligns with its empirical accuracy, higher predictive uncertainty signals harder instances. To ground Urank\mathcal{U}_{\mathrm{rank}} on a calibrated source of uncertainty, we adopt the variational Bayesian last layer (VBLL) [19] as the prediction head. Trained with a variational evidence lower bound, VBLL achieves lower expected calibration error than deterministic baselines across regression and classification tasks [19]. Its predictive variance σ2\sigma^2 quantifies epistemic uncertainty [24]: a larger value reflects insufficient evidence for a confident prediction and therefore serves as a proxy for instance complexity.

The list-wise extension preserves this probabilistic grounding. The calibration argument above concerns the predictive distribution over a single outcome, while sequential recommendation operates on ordered lists. Urank bridges this gap through two probabilistic constructions. The pairwise outranking probability pi≻jp_{i\succ j} (Eq. (2)) is the closed-form solution of the Bayesian Thurstone model [42] under Gaussian distribution. Aggregating these probabilities through the negative log-likelihood (Eq. (3)) induces a coherent measure of list-wise ranking confidence, approaching its minimum when all top-KK pairs are confidently separated. The resulting Urank\mathcal{U}_{\mathrm{rank}} is therefore a functional of the predictive distribution [24], reflecting the structural stability of the predicted ordering as a list-wise complexity signal. We empirically verify this argument in § 4.4.

Self-Supervised Threshold Prediction Training

While ranking uncertainty provides a reliable measure of instance complexity, computing it requires a full forward pass including the final prediction layer. To enable intelligent resource allocation, we need to anticipate the ranking uncertainty early in inference. Specifically, we design a lightweight threshold predictor gϕ:Rd→[0,1]g_{\phi}: \mathbb{R}^{d} \rightarrow[0,1] to predict the ideal slimming rate. The predictor takes a context vector c∈Rd\mathbf{c} \in\mathbb{R}^{d} as a user embedding, which we obtain by extracting the last valid token representation from early-layer representations (details in § 3.4 and § 3.5). The threshold predictor is implemented with a simple two-layer MLP:

τ=gϕ(c),(4)\tau= g_{\phi}(\mathbf{c}), \tag*{(4)}

where τ∈[0,1]\tau\in[0,1] directly controls structure customization, while a linearly transformed variant τ′\tau' governs sequence customization (detailed in § 3.4 and § 3.5).

To early capture uncertainty, we establish a self-supervised learning objective. We compute the ground-truth ranking uncertainty Urank\mathcal{U}_{\mathrm{rank}} and normalize it using the current mini-batch statistics:

U~=Urank−E[Urank]std⁡(Urank)+ϵ.(5)\widetilde{\mathcal{U}} = \frac{\mathcal{U}_{\mathrm{rank}}-\mathbb{E}[\mathcal{U}_{\mathrm{rank}}]}{\operatorname{std}(\mathcal{U}_{\mathrm{rank}})+\epsilon}. \tag*{(5)}

We then map the normalized uncertainty to a target threshold τ∗\tau^* as our ideal slimming rate through a sigmoid transformation:

τ∗=σ(−U~+bcenter),(6)\tau^* = \sigma\left(-\widetilde{\mathcal{U}}+b_{\mathrm{center}}\right), \tag*{(6)}

The uncertainty-derived target τ∗\tau^* is detached during predictor training, where bcenter=log⁡(r01−r0)b_{\mathrm{center}}=\log\left(\frac{r_0}{1-r_0}\right) is designed to center the threshold distribution around a target pruning ratio r0r_0. The negative sign ensures that higher uncertainty (difficult instances) leads to lower ratios, which results in less aggressive pruning and greater model capacity. During training, the threshold predictor is guided by the pseudo-label τ∗\tau^* using the mean squared error:

Lthreshold=Es∼D[(τ−τ∗)2].(7)\mathcal{L}_{\mathrm{threshold}}=\mathbb{E}_{\mathbf{s}\sim\mathcal{D}}\left[(\tau-\tau^*)^2\right]. \tag*{(7)}

This self-supervised approach allows the predictor to learn the mapping from early representations to appropriate resource allocation without requiring external labels or manual annotations.

Interest-Compatible Subnet Identification

Once the instance-specific resource budget (compression ratio τ\tau) is determined by the uncertainty predictor, the next challenge is deciding which specific parameters to retain. Existing methods often rely on external auxiliary models to extract user features [11], which ironically introduces additional computational overhead. To address this, RACE adopts an Intrinsic Feature Reuse strategy. We posit that the model’s intermediate representations already contain useful information about user intent to guide its own structural adaptation, eliminating the need for heavy external extractors.

Formally, let H(l)∈RT×d\mathbf{H}^{(l)} \in\mathbb{R}^{T \times d} denote the hidden states at the ll-th Transformer layer. We extract the representation of the last valid item in the sequence as the dynamic user embedding c(l)\mathbf{c}^{(l)}, which serves as the context for decision-making. To determine the importance of each channel in the Feed-Forward Network (FFN), we employ a lightweight Sensitivity Extractor. This module, acting as a hypernetwork-style scorer, projects the user embedding c(l)\mathbf{c}^{(l)} into channel-wise sensitivity scores s(l)∈RdFF\mathbf{s}^{(l)} \in\mathbb{R}^{d_{\mathrm{FF}}}:

s(l)=σ(Wsc(l)+bs),(8)\mathbf{s}^{(l)} = \sigma\left(\mathbf{W}_{s}\mathbf{c}^{(l)}+\mathbf{b}_{s}\right), \tag*{(8)}

where Ws\mathbf{W}_{s} and bs\mathbf{b}_{s} are learnable parameters, and σ\sigma is the sigmoid function. These scores represent the potential contribution of each neuron to the current user’s inference.

To construct the channel mask m(l)\mathbf{m}^{(l)}, we utilize the predicted channel pruning ratio τ\tau to perform dynamic truncation. Specifically, we identify the threshold value δ\delta at the τ\tau-quantile of s(l)\mathbf{s}^{(l)} and generate the mask as follows:

mi(l)=I(si(l)≥δ),(9)\mathbf{m}_{i}^{(l)}=\mathbb{I}\left(s_{i}^{(l)}\ge\delta\right), \tag*{(9)}

where I(⋅)\mathbb{I}(\cdot) is the indicator function. Thus, τ\tau is the pruned-channel fraction, and the default τ=0.8\tau=0.8 prunes approximately 80% of the channels. During training, we use a Straight-Through Estimator (STE): the hard mask is detached, while gradients are passed through the continuous channel scores. By applying this mask with the FFN weights, RACE dynamically carves out interest-compatible subnets.

Progressive Sequence Filtering

Guided by the predicted ratio τ′\tau', RACE implements a dynamic data-slimming strategy to identify crucial items. Drawing inspiration from token pruning in natural language processing [16, 26, 52], we propose Progressive Sequence Filtering to eliminate redundancy of user interaction history block by block, ensuring the computational budget is concentrated on the most informative items.

In Transformer-based architectures, the self-attention map A(l)∈RH×T×T\mathbf{A}^{(l)} \in\mathbb{R}^{H \times T \times T} naturally encodes the contribution of each item. However, in sequential recommendation, standard attention scores are subject to causal masking and can be volatile across layers. To capture the relative significance for each interaction, we design a cumulative importance metric based on the head-averaged attention map Aˉ(l)=1H∑A(l)\bar{\mathbf{A}}^{(l)} = \frac{1}{H}\sum\mathbf{A}^{(l)}. As illustrated in Figure 1(c), for each item t∈{1,…,T}t \in\{1,\ldots,T\}, its importance score sts_t is formulated as the normalized product of its weighted contributions across the sequence:

st=s~t∑j=1Ts~j+ϵ,wheres~t=∏j=tT(Aˉj,t(l)⋅η(j)).(10)s_t = \frac{\tilde{s}_t}{\sum_{j=1}^{T}\tilde{s}_j+\epsilon}, \qquad\text{where}\quad\tilde{s}_t = \prod_{j=t}^{T}\left(\bar{A}^{(l)}_{j,t}\cdot\eta(j)\right). \tag*{(10)}

Here, η(j)=j\eta(j)=j serves as a position-aware factor that linearly scales attention weights by the query position jj, effectively prioritizing later queries that encode more recent user interests. The product operator (Π\Pi) functions as a cumulative filter. This approach accumulates attention scores from later, more informative queries, effectively prioritizing items that consistently contribute to the user’s current interests.

Using the uncertainty-driven layer-wise pruning ratio τ′\tau', block ll retains nlkeep=⌈∣S(l)∣(1−τ′)⌉n_l^{\mathrm{keep}}=\lceil|S^{(l)}|(1-\tau')\rceil items:

S(l+1)=TopK⁡(S(l),s,nlkeep).(11)S^{(l+1)}=\operatorname{TopK}\left(S^{(l)},s,n_l^{\mathrm{keep}}\right). \tag*{(11)}

In the training phase, we employ a "soft" filtering approach by masking unselected tokens to maintain batch stability while simulating the pruning effect. By contrast, during inference, unselected tokens are dropped to achieve immediate FLOPs reduction.

Loss Function

Our training objective jointly optimizes recommendation accuracy, resource efficiency (§ 3.4 and § 3.5), and uncertainty-aware adaptation (§ 3.2). The complete loss function consists of three components:

Ltotal=Lrec+λ1Lsparsity+λ2Lthreshold.(12)\mathcal{L}_{\mathrm{total}}=\mathcal{L}_{\mathrm{rec}}+\lambda_1\mathcal{L}_{\mathrm{sparsity}}+\lambda_2\mathcal{L}_{\mathrm{threshold}}. \tag*{(12)}

where Lrec\mathcal{L}_{\mathrm{rec}} is the recommendation loss, combining cross-entropy and KL divergence for Bayesian inference. Lsparsity=(ρ−rtarget)2\mathcal{L}_{\mathrm{sparsity}}=(\rho-r_{\mathrm{target}})^2 penalizes deviation from average sparsity ratio rtargetr_{\mathrm{target}}, where ρ\rho is the average channel slimming ratio. Lthreshold\mathcal{L}_{\mathrm{threshold}} (Eq. (7)) trains the threshold predictor to anticipate the ideal pruning ratio. λ1\lambda_1 and λ2\lambda_2 are scaling coefficients used to balance the trade-off between accuracy, average sparsity, and dynamic resource allocation.

EXPERIMENTS

We conduct extensive experiments on four public datasets and answer the following research questions:

  • RQ1: Does RACE outperform existing methods in accuracy while simultaneously reducing computational cost?

  • RQ2: Is each component of RACE necessary and superior to existing alternatives?

  • RQ3: Does list-wise ranking-uncertainty reliably reflect instance complexity and effectively drive RACE’s adaptive allocation?

  • RQ4: Does RACE achieve stable and robust training behavior?

  • RQ5: Does RACE achieve efficient, adaptive inference under varying resource constraints?

Experimental Setup

Datasets.

The experiments are conducted on four datasets: MovieLens-100k1, Douban [62], Amazon-CD2 and Amazon-Game. Following the 10-core setting, we exclude users and items with fewer than 10 interactions. For each user, the most recent 10% of interactions form the test set and the rest are used for training.

Baselines.

We adopt SASRec [23] and BSARRec [38] as backbones and compare RACE with nine baselines: (1) Full-precision Methods: Normal and Finetune. (2) Compression-based Methods:

  • SNIP [28]: Prunes connections by saliency before training.

  • GEM [39]: Finds high-performance subnets at initialization.

  • APG [48]: Decomposes weights via low-rank approximation.

  • DIET [11]: Customizes model parameters for individual devices through multi-level sensitivity analysis.

  • ATP-LLAVA [53]: Prunes redundant tokens dynamically.

  • PerSRec [58]: Compresses histories into a fixed number of personalized learnable tokens, keeping recent interactions.

  • CAUSE [31]: Merges histories into category-based tokens.

Evaluation Metrics.

We evaluate recommendation quality with Hit@5 (Hit) and NDCG@5 (NDCG) under full ranking over the entire item set to avoid candidate-sampling bias, and measure computational cost by backbone inference FLOPs (MFLOPs).

Implementation Detail.

For all datasets, we set the maximum sequence length to 20 with zero padding, the item embedding dimension to 64, and four Transformer blocks for both backbones. The loss coefficients λ1\lambda_1 and λ2\lambda_2 (§ 3.6) are set to 1 and 0.5, respectively. For compression-based methods, we tune their hyperparameters to reach sparsity similar to ours without much degradation.

Overall Performance(RQ1)

Table 1 presents the overall performance comparison across four datasets and two backbones. From the results, we can observe that Obs.1: RACE consistently achieves the best performance in both accuracy and efficiency across all settings. Averaged over the four datasets, RACE improves NDCG and Hit by 40.87% and 35.12% on SASRec while reducing FLOPs by 2.37×. On BSARRec, the corresponding average improvements are 47.73% and 41.77%, together with a 2.49× FLOPs reduction.

ModelMethodMovieLens-100k NDCG ↑MovieLens-100k Hit ↑MovieLens-100k FLOPs ↓Douban NDCG ↑Douban Hit ↑Douban FLOPs ↓Amazon-CD NDCG ↑Amazon-CD Hit ↑Amazon-CD FLOPs ↓Amazon-Game NDCG ↑Amazon-Game Hit ↑Amazon-Game FLOPs ↓
SASRecNormal0.02990.04711.84320.01410.01931.84320.02500.02951.84320.03340.04181.8432
Finetune0.02680.04421.84320.01310.01741.84320.02510.02971.84320.03380.04241.8432
SNIP0.02850.04640.83520.01090.01530.83520.02400.02800.83520.03310.04010.8352
GEM0.03300.05160.83520.01440.01990.83520.02630.03120.83520.03430.04220.8352
APG0.03380.05360.89470.02240.03050.89470.02670.03120.89470.03850.04680.8947
DIET0.03280.05211.01630.00830.01130.81670.02540.02870.90100.03970.04810.8925
ATP-LLAVA0.02400.03770.90100.00860.01340.94910.02540.02981.01430.03110.03730.9815
PerSRec0.02590.04110.87040.00670.01030.87040.01060.01500.87040.01750.02560.8704
CAUSE0.02770.04420.87040.00470.00700.87040.02480.02910.87040.03160.04030.8704
RACE0.04270.06460.80250.02610.03370.79280.02880.03290.75490.04020.04900.7660
Improv.42.81%37.15%2.30×85.11%74.61%2.32×15.20%11.50%2.44×20.36%17.22%2.41×
BSARecNormal0.02560.04094.13700.02060.02754.13700.03680.04314.13700.03250.04044.1370
Finetune0.03050.04754.13700.01640.02164.13700.03830.04514.13700.03120.03954.1370
SNIP0.03220.05091.74540.03050.03911.74540.04180.04921.74540.03350.04291.7454
GEM0.02980.04651.74540.02190.03021.74540.03880.04551.74540.03070.03951.7454
APG0.03580.05642.16650.02820.03692.16650.04090.04812.16650.03270.04202.1665
DIET0.03460.05402.38450.02030.02692.34900.04190.04822.46600.03390.04241.8703
ATP-LLAVA0.02870.04702.82430.01110.01632.67700.00830.01052.51610.01080.01562.2927
PerSRec0.02790.04422.01730.01560.02162.01730.03490.03962.01730.03280.03952.0173
CAUSE0.02660.04212.01730.01350.01922.01730.03810.04382.01730.03180.03952.0173
RACE0.03960.06001.71480.03270.04181.64730.04820.05421.64320.04760.05761.6424
Improv.54.69%46.70%2.41×58.74%52.00%2.51×31.01%25.82%2.52×46.46%42.57%2.52×

Table 1. Overall performance comparison of RACE and competitive baselines across four datasets and two representative backbones. For metrics NDCG@KNDCG@K and Hit@KHit@K, K=5K = 5 is used by default. ↑ and ↓ denote that larger and smaller values are preferred, respectively. Bold font indicates the best performance in each column.

Among the baselines, finetuning with limited data does not guarantee performance improvement. SNIP and GEM apply uniform masks across all instances and fail to adapt to instance interests and complexity. APG and DIET customize the model side through low-rank decomposition and sensitivity analysis, respectively, and ATP-LLAVA dynamically prunes tokens, yet these single-side methods show unstable accuracy across datasets. PerSRec and CAUSE compress user histories into learnable or category-based tokens but retain fixed compression budgets. RACE also improves over the strongest competing baseline in representative settings: on Douban with SASRec, it achieves 0.0261 NDCG, 16.52% above APG (0.0224); on Amazon-CD with BSARec, it achieves 0.0482 NDCG, 15.04% above DIET (0.0419), while using fewer FLOPs than CAUSE (1.6432 vs. 2.0173 MFLOPs).

ModelAblated ModelMovieLens-100k NDCG ↑MovieLens-100k Hit ↑Douban NDCG ↑Douban Hit ↑
SASRecw/o uncertainty0.02990.04710.01410.0193
w/o customization0.02630.04470.01390.0193
w/o sequence filtering0.03880.05920.02290.0296
w/o subset identification0.02960.04920.01040.0171
w/o dynamic ratio0.03760.05840.02500.0321
RACE0.04270.06460.02610.0337
BSARecw/o uncertainty0.02560.04090.02060.0275
w/o customization0.02750.04410.00950.0150
w/o sequence filtering0.03100.05070.02290.0304
w/o subset identification0.02680.04430.00940.0141
w/o dynamic ratio0.03260.05200.02330.0315
RACE0.03960.06000.03270.0418

Table 2. Ablation experiment.

Ablation Study (RQ2)

Table 2 reports the ablation results, and Table 3 and Table 4 compare two key components of RACE with alternative designs.

Uncertainty DesignSASRecBSARec
NDCG ↑Hit ↑NDCG ↑Hit ↑
MC Dropout0.02160.02920.02840.0366
Ensemble0.01970.02660.02610.0344
Mean Variance0.02480.03170.03100.0415
Entropy0.02270.03110.02960.0389
Pairwise Margin0.02280.02890.03230.0407
LIDU0.02590.03370.02890.0378
RACE0.02610.03370.03270.0418

Table 3. Comparison of alternative uncertainty designs for guiding dynamic thresholds on Douban.

ModelMethodDoubanAmazon-Game
NDCG ↑Hit ↑NDCG ↑Hit ↑
SASRecMean0.02380.03120.03930.0471
SASRecMasked Mean0.02410.03180.03900.0474
SASRecRACE0.02610.03370.04020.0490
BSARecMean0.03100.03950.04720.0567
BSARecMasked Mean0.03070.03970.04770.0568
BSARecRACE0.03270.04180.04760.0576

Table 4. Comparison of different sequence-filtering methods on Douban and Amazon-Game datasets.

Component Ablation

Obs.2: Every component of RACE plays a meaningful and complementary role.

  • w/o Uncertainty collapses RACE to the unified backbone and eliminates all gains.

  • w/o Customization replaces the backbone’s prediction layer with a variational Bayesian layer for uncertainty estimation, showing that uncertainty alone cannot model user interests accurately and even degrades performance.

  • w/o Sequence Filtering keeps only interest-aware subnet identification, showing that mitigating model redundancy contributes meaningfully, with a 29.8% NDCG improvement on MovieLens-100K with SASRec.

  • w/o Subnet Identification retains sequence filtering but disables channel-wise subnet customization, showing that reducing data redundancy alone sometimes degrades performance, e.g., on Douban.

  • w/o Dynamic Ratio applies both customizations with a fixed pruning ratio, showing that jointly alleviating model and data redundancy benefits both accuracy and efficiency and outperforms the full-precision model.

  • RACE further allocates model and input size dynamically under uncertainty guidance. Its gap over w/o Dynamic Ratio (e.g., over 40% on Douban/BSARec) validates dynamic resource allocation based on instance complexity.

Comparison with Alternative Designs

Obs.3: RACE’s design choices outperform alternative designs in both uncertainty estimation and sequence filtering.

Comparison with alternative uncertainty designs. Table 3 compares two levels of uncertainty design: predictive uncertainty estimation and ranking-uncertainty formulation. At the first level, we replace the variational Bayesian last layer (VBLL) [19] in RACE’s prediction head with MC Dropout [13] and Ensemble [27], both restricted to the prediction layer for a compute-aligned comparison. VBLL produces predictive means and variances in one pass without repeated stochastic passes or multiple heads, and yields NDCG@5 of 0.0261/0.0327 versus 0.0216/0.0284 for the stronger alternative on SASRec/BSARec. At the second level, with VBLL fixed, Mean Variance and Entropy provide point-wise summaries [7], and Pairwise Margin and LIDU [30] add local ranking information, whereas our list-wise ranking uncertainty models the distributional stability of the entire top-KK list by considering interdependencies among candidates. RACE achieves the best accuracy on both backbones, supporting VBLL as an efficient technical choice and, more importantly, validating our list-wise uncertainty design for dynamic threshold prediction.

Comparison with other item-importance metrics. Beyond uncertainty estimation, Table 4 compares item-importance metrics for sequence filtering. The global average of attention scores across heads (Mean) [53], common in prior work, ignores the causal structure of sequential recommendation and causes an importance mismatch. Applying a causal mask to the average (Masked Mean) still suffers from positional bias, where earlier items receive disproportionate attention regardless of their relevance. RACE’s cumulative and position-aware metric prioritizes recent and semantically core interactions, achieving the best performance on seven of the eight reported metrics and remaining comparable on the other. For instance, on Douban with SASRec, RACE raises NDCG to 0.0261, compared with 0.0238 for Mean and 0.0241 for Masked Mean.

Uncertainty-Driven Resource Allocation Analysis (RQ3)

Figure 4 examines ranking uncertainty from two perspectives: its correlation with recommendation accuracy across uncertainty deciles (before or after customization), and the distribution of predicted pruning ratios across uncertainty groups.

Uncertainty analysis

Figure 4. Uncertainty Analysis: (a) Performance (before or after customization) vs. Uncertainty (b) Predicted Channel Pruning Ratio vs. Uncertainty.

Obs.4: List-wise ranking-uncertainty serves as an effective proxy for instance complexity. In Figure 4(a), we visualize the relationship between ranking uncertainty and recommendation performance across ten uncertainty deciles on CD dataset with SASRec backbone. Observing the gray curve (before customization), we find a clear negative correlation: instances with lower uncertainty (D1 to D3) consistently achieve higher NDCG scores, while high-uncertainty instances (D8 to D10) exhibit clearly degraded performance. This empirically validates the theoretical argument in Section 3.2.2 that ranking uncertainty serves as a principled, data-driven proxy for instance complexity.

Obs.5: RACE intelligently allocates resources based on uncertainty-aware threshold prediction and benefits instances of varying complexity through uncertainty-guided dual-space customization. Comparing performance before (gray line) and after (orange line) customization in Figure 4(a) reveals that RACE improves recommendation accuracy for most uncertainty groups and reduces performance gaps between them. High-uncertainty instances show the largest observed gains from additional capacity allocation, while low-uncertainty instances maintain comparable performance with reduced computation costs, demonstrating that RACE successfully addresses the Capacity-Complexity Mismatch through uncertainty-guided multi-level customization.

To further verify whether RACE learns to allocate resources intelligently, Figure 4(b) shows the predicted channel pruning ratios across five uncertainty percentile groups (0–20%, 20–40%, 40–60%, 60–80%, 80–100%), where boxes span the 25th to 75th percentiles, lines mark the medians, and whiskers extend to the 10th and 90th percentiles. The results reveal a clear downward trend: the negative correlation (Pearson’s r=−0.319r = -0.319, Spearman’s ρ=−0.496\rho= -0.496) and the fitted trend line confirm that RACE automatically learns to preserve more capacity for complex instances while aggressively compressing simpler ones, validating the effectiveness of our uncertainty-guided resource allocation.

Training Robustness Analysis (RQ4)

Obs.6: RACE exhibits stable training dynamics and remains robust across hyperparameter choices.

Training stability analysis. Figure 5 plots RACE’s training curves on Amazon-CD over five random seeds with SASRec and BSARec. The tight ±1 std bands and low coefficients of variation in the converged region (1.18% and 0.50%, respectively) confirm stable optimization across random initializations.

Training Curve of RACE over Five Seeds

Figure 5. Training Curve of RACE over Five Seeds.

Pruning ratio sensitivity analysis. Figures 6 and 7 examine the layer-wise sequence and channel pruning ratios, respectively. Sequence pruning peaks at 0.4 on SASRec, while BSARec varies only mildly from 0.2 to 0.5. Channel pruning favors the high-ratio regime: 0.8 performs best on SASRec, and BSARec remains strong from 0.7 to 0.9. We therefore use 0.4 and 0.8 as shared defaults to maintain strong performance across both backbones while retaining substantial pruning.

Sensitivity Analysis of Sequence Pruning Ratio

Figure 6. Sensitivity Analysis of Sequence Pruning Ratio.

Sensitivity Analysis of Channel Pruning Ratio

Figure 7. Sensitivity Analysis of Channel Pruning Ratio.

Loss-coefficient sensitivity analysis. On BSARec/Amazon-CD, the default (λ1=1\lambda_1 = 1, λ2=0.5\lambda_2 = 0.5) yields the best observed result; NDCG varies by at most 2.1%/6.4% across the evaluated λ1/λ2\lambda_1/\lambda_2 ranges, respectively.

Inference Efficiency and Adaptivity Analysis (RQ5)

We evaluate inference efficiency under typical resource budgets and adaptivity to test-time constraints that differ from training.

Efficiency Analysis

To evaluate efficiency under resource constraints, we simulate local reranking with a pool of 50 candidates, prefetched from the cloud, following established on-device settings [11, 14, 15]. On SASRec, Table 5 reports backbone and auxiliary FLOPs together with latency profiled on a 4-core CPU.

MethodBackbone FLOPs ↓Other FLOPs ↓Latency ↓
Finetune1.84323.68649.732
DIET0.90103.059710.518
RACE0.75490.04905.780

Table 5. Efficiency analysis on the SASRec backbone.

Obs.7: RACE achieves efficient inference with negligible auxiliary cost through intrinsic feature reuse. Compared with Finetune, RACE reduces backbone FLOPs from 1.8432 to 0.7549 MFLOPs, a 59.0% reduction. Its 0.0490-MFLOP auxiliary overhead is 62× lower than DIET’s (3.0597 MFLOPs) and 75× lower than Finetune’s backpropagation cost. At 5.780 ms per CPU inference, RACE reduces latency by 40.6% and 45.0% relative to Finetune and DIET, respectively.

Adaptivity Analysis

Table 6 evaluates RACE on Douban and CD with SASRec under test-time constraints different from training.

DatasetTest-time ConstraintNDCG ↑Hit ↑FLOPs ↓
DoubanStricter than training0.02590.03370.6266
DoubanSame as training0.02610.03370.7928
DoubanLooser than training0.02620.03411.0408
CDStricter than training0.02870.03280.6718
CDSame as training0.02880.03290.7549
CDLooser than training0.02880.03300.8567

Table 6. Adaptivity analysis under varying test-time constraints on the SASRec backbone.

Obs.8: RACE gracefully trades accuracy and compute under varying inference constraints. Under stricter constraints than training, RACE incurs only 0.8% / 0.3% NDCG drops on Douban and CD while reducing FLOPs by an additional 21.0% / 11.0%. Under looser constraints, the extra budget yields consistent NDCG and Hit improvements across both datasets. Together, these results demonstrate that RACE learns a flexible resource allocation policy that can adapt to varying inference constraints without retraining.

Conclusion

In this paper, we identify the Capacity-Complexity Mismatch in efficient sequential recommendation and propose RACE, which moves resource-constrained local inference from static-capacity compression to dynamic, uncertainty-aware adaptation. Guided by self-supervised list-wise ranking uncertainty as a proactive resource signal, RACE performs multi-granularity slimming across parameter and data spaces via hypernetwork-style channel customization and importance-aware progressive filtering, with low overhead from intrinsic feature reuse. Extensive experiments show that RACE balances recommendation accuracy and inference efficiency across diverse resource-constrained scenarios. Beyond on-device compression, this label-free self-assessment generalizes to broader sequential recommendation tasks such as performance estimation and model refresh scheduling. In future work, we plan to extend this paradigm to generative recommenders, where ranking uncertainty could guide adaptive decoding depth and selective beam expansion.

Ethical Considerations

RACE does not require additional data collection beyond the interaction histories used by the underlying recommender and does not alter its recommendation objective. Practical deployments should follow the same data-governance, access-control, and quality-evaluation practices as the underlying system.

Supplementary Material

Loss Coefficient Sensitivity Analysis

We additionally examine the sensitivity of the two coefficients in the training objective (Eq. (12)) on BSARec with the Amazon-CD dataset. Figure S1 reports NDCG@5 and Hit@5 while varying one coefficient and fixing the other at its default value.

Sensitivity plots of NDCG@5 and Hit@5 to loss coefficients

Figure S1. Sensitivity of NDCG@5 and Hit@5 to loss coefficients λ1\lambda_1 and λ2\lambda_2 on BSARec with Amazon-CD. Stars mark the default setting used in the main experiments.

Among the evaluated settings, the best observed configuration is λ1=1\lambda_1 = 1 and λ2=0.5\lambda_2 = 0.5. NDCG@5 varies by at most 2.1% as λ1\lambda_1 ranges from 0.1 to 10 and by at most 6.4% as λ2\lambda_2 ranges from 0.01 to 5. The smaller variation for λ1\lambda_1 and the moderate variation for λ2\lambda_2 show that the default is not a single sharply tuned point, while also indicating that performance is more sensitive to λ2\lambda_2. Overall, the results support robustness across the evaluated coefficient ranges without implying complete invariance to their choice.

Pseudocode

We place the pseudocode of RACE in Algorithm S1. RACE can be summarized as six modules: Modules 1–5 form the inference pipeline that produces the Top-kk list under instance-specific channel and sequence pruning, while Module 6 runs only during training and supervises the threshold predictor (Module 2) via the pseudo-label τ∗\tau^* derived from Urank\mathcal{U}_{\mathrm{rank}}. Together, the six modules realize an instance-specific, uncertainty-guided slimming pipeline that adaptively allocates model and input capacity to each instance. Here, τ\tau and τ′\tau' denote channel and sequence pruning ratios, respectively, and larger values indicate more aggressive pruning.

RACE anticipates the list-wise uncertainty from early-layer features and decides the pruning ratios τ\tau and τ′\tau' accordingly in Module 2, determining how aggressively each instance is slimmed in subsequent modules. For a hard instance with high anticipated ranking uncertainty, Module 2 outputs smaller ratios, leading Module 3 to retain more FFN channels in each Transformer block and Module 4 to keep more interaction items as the sequence propagates across blocks. Conversely, for an easy instance, larger thresholds drive more aggressive channel deactivation and sequence filtering, reducing inference cost more substantially while still delivering personalized and accurate recommendations.

Algorithm S1 Pseudocode of RACE

Module 1: ▷ Early-Layer Feature Encoding

Function: Encode the historical sequence through the 0-th Transformer block F(0)\mathcal{F}^{(0)} and repurpose the last valid token representation as a dynamic user embedding c\mathbf{c}. Input: Historical Sequence Su\mathcal{S}^{u} Output: Early-layer Features H(0)\mathbf{H}^{(0)}, Dynamic User Embedding c\mathbf{c}

Module 2: ▷ Self-Supervised Threshold Prediction

Function: Anticipate instance-specific pruning ratios from the dynamic user embedding via a lightweight predictor gϕg_{\phi}, supervised by the ideal threshold τ∗\tau^* from Module 6. Input: Dynamic User Embedding c\mathbf{c} Output: Channel Pruning Ratio τ\tau, Sequence Pruning Ratio τ′\tau'

Module 3: ▷ Interest-Compatible Subnet Identification

Function: Analyze FFN channel sensitivities via the hypernetwork-style Sensitivity Extractor and carve out an instance-specific subnet according to channel pruning ratio τ\tau. Input: Hidden States H(l)\mathbf{H}^{(l)}, Channel Pruning Ratio τ\tau Output: Channel Mask m(l)\mathbf{m}^{(l)}

Module 4: ▷ Progressive Sequence Filtering

Function: Compute the cumulative importance score sts_t block by block and retain the top-nlkeepn_l^{\mathrm{keep}} items, where nlkeep=∣S(l)∣(1−τ′)n_l^{\mathrm{keep}} = \lvert\mathcal{S}^{(l)}\rvert(1-\tau'). Input: Attention Map A(l)\mathbf{A}^{(l)}, Sequence Pruning Ratio τ′\tau' Output: Retained Token Set S(l+1)\mathcal{S}^{(l+1)}

Module 5: ▷ Capacity-Adaptive Inference

Function: Achieve capacity-adaptive inference for model and data, skipping pruned channels and filtered sequence block by block to generate the recommendation list. Input: Hidden States H(l)\mathbf{H}^{(l)}, Channel Mask m(l)\mathbf{m}^{(l)}, Retained Item Set S(l+1)\mathcal{S}^{(l+1)} Output: Top-kk Recommendation List

Module 6: ▷ List-Wise Ranking Uncertainty (training only)

Function: Quantify ranking confidence via list-wise ranking uncertainty calculations over top-KK predictions and derive the pseudo-label τ∗\tau^* to supervise Module 2. Input: Top-KK Predictive Mean μK\boldsymbol{\mu}_K, Variance σK2\boldsymbol{\sigma}_K^2 Output: Ranking Uncertainty Urank\mathcal{U}_{\mathrm{rank}}, Ideal Channel Pruning Ratio τ∗\tau^*

References

  1. [1]Yue Bai, Huan Wang, Xu Ma, Yitian Zhang, Zhiqiang Tao, and Yun Fu. 2022. Parameter-efficient masking networks. Advances in Neural Information Processing Systems 35 (2022), 10217–10229.arxiv.org/abs/2210.06699
  2. [2]Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning. 41–48.DOI
  3. [3]Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. 2015. Weight uncertainty in neural network. In International conference on machine learning. PMLR, 1613–1622.
  4. [4]Renqin Cai, Jibang Wu, Aidan San, Chong Wang, and Hongning Wang. 2021. Category-aware collaborative sequential recommendation. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 388–397.DOI
  5. [5]Shuyu Chen, Yang Zhang, Jiliang Tang, Dawei Yin, Yi Chang, Mark A Hasegawa-Johnson, and Thomas Huang. 2017. Streaming recommender systems. In Proceedings of the 26th international conference on world wide web. 381–389.arxiv.org/abs/1607.06182
  6. [6]Zhourong Chen, Yang Li, Samy Bengio, and Si Si. 2019. You look twice: Gaternet for dynamic filter selection in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9172–9180.arxiv.org/abs/1811.11205
  7. [7]Victor Coscrato and Derek Bridge. 2023. Estimating and Evaluating the Un-certainty of Ranking Predictions and Top-n Recommendations in Recommender Systems. ACM Trans. Recomm. Syst. 1, 2, Article 7 (April 2023), 34 pages. doi:10.1145/3584021DOI
  8. [8]Sunhao Dai, Ningyuan Shao, Jieming Zhu, Xiao Zhang, Zhenhua Dong, Jun Xu, Quanyu Dai, and Ji-Rong Wen. 2024. Modeling user attention in music recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 761–774.DOI
  9. [9]Jin Du, Zesheng Ye, Bin Guo, Zhiwen Yu, and Lina Yao. 2023. Indp: Interest dynamics modeling using generative neural processes for sequential recommendation. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining. 481–489.arxiv.org/abs/2208.04600
  10. [10]Kairui Fu, Changfa Wu, Kun Yuan, Binbin Cao, Dunxian Huang, Yuliang Yan, Jun-Jun Zheng, Jianning Zhang, Silu Zhou, Jian Wu, and Kun Kuang. 2026. RankGR: Rank-Enhanced Generative Retrieval with Listwise Direct Preference Optimization in Recommendation. arXiv preprint arXiv:2602.08575 (2026).arxiv.org/abs/2602.08575
  11. [11]Kairui Fu, Shengyu Zhang, Zheqi Lv, Jingyuan Chen, and Jiwei Li. 2024. DIET: Customized slimming for incompatible networks in sequential recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 816–826.arxiv.org/abs/2406.08804
  12. [12]Kairui Fu, Tao Zhang, Shuwen Xiao, Ziyang Wang, Xinming Zhang, Chenchi Zhang, Yuliang Jan, Junjun Zheng, Xiangheng Kong, Shengyu Zhang, Kun Kuang, and Yuning Jiang. 2026. FORGE: Forming Semantic Identifiers for Generative Retrieval in Industrial Datasets. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 8. 8858–8869.
  13. [13]Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning. PMLR, 1050–1059.arxiv.org/abs/1506.02142
  14. [14]Xudong Gong, Qinlin Feng, Yuan Zhang, Jiangling Qin, Weijie Ding, Biao Li, Peng Jiang, and Kun Gai. 2022. Real-time short video recommendation on mobile devices. In Proceedings of the 31st ACM international conference on information & knowledge management. 3103–3112.
  15. [15]Yu Gong, Ziwen Jiang, Yufei Feng, Binbin Hu, Kaiqi Zhao, Qingwen Liu, and Wenwu Ou. 2020. EdgeRec: recommender system on edge in Mobile Taobao. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 2477–2484.arxiv.org/abs/2005.08416
  16. [16]Saurabh Goyal, Anamrita Roy Choudhury, Saurabh Raje, Venkatesan Chakravarthy, Yogish Sabharwal, and Ashish Verma. 2020. Power-bert: Accelerating bert inference via progressive word-vector elimination. In International Conference on Machine Learning. PMLR, 3690–3699.
  17. [17]Yue Guan, Zhengyi Li, Jingwen Leng, Zhouhan Lin, and Minyi Guo. 2022. Trans-kimmer: Transformer learns to layer-wise skim. arXiv preprint arXiv:2205.07324 (2022).arxiv.org/abs/2205.07324
  18. [18]Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning. PMLR, 1321–1330.arxiv.org/abs/1706.04599
  19. [19]James Harrison, John Willes, and Jasper Snoek. 2024. Variational Bayesian last layers. arXiv preprint arXiv:2404.11599 (2024).arxiv.org/abs/2404.11599
  20. [20]Maria Heuss, Daniel Cohen, Masoud Mansoury, Maarten de Rijke, and Carsten Eickhoff. 2023. Predictive uncertainty-based bias mitigation in ranking. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 762–772.DOI
  21. [21]Eyke Hüllermeier and Willem Waegeman. 2021. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Machine learning 110, 3 (2021), 457–506.DOI
  22. [22]Hyunsik Jeon, Se-eun Yoon, and Julian McAuley. 2024. Calibration-disentangled learning and relevance-prioritized reranking for calibrated sequential recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 973–982.DOI
  23. [23]Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM). IEEE, 197–206.
  24. [24]Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems 30 (2017).arxiv.org/abs/1703.04977
  25. [25]Wazir Zada Khan, Ejaz Ahmed, Saqib Hakak, Ibrar Yaqoob, and Arif Ahmed. 2019. Edge computing: A survey. Future Generation Computer Systems 97 (2019), 219–235.
  26. [26]Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. 2022. Learned token pruning for transformers. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 784–794.arxiv.org/abs/2107.00910
  27. [27]Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30 (2017).arxiv.org/abs/1612.01474
  28. [28]Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. 2018. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340 (2018).arxiv.org/abs/1810.02340
  29. [29]Chenyi Lei, Yong Liu, Lingzi Zhang, Guoxin Wang, Haihong Tang, Houqiang Li, and Chunyan Miao. 2021. A sequential multi-modal information transfer network for e-commerce micro-video recommendations. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 3161–3171.DOI
  30. [30]Jiayu Li, Ziyi Ye, Guohao Jian, Zhiqiang Guo, Weizhi Ma, Qingyao Ai, and Min Zhang. 2025. Are Recommenders Self-Aware? Label-Free Recommendation Performance Estimation via Model Uncertainty. arXiv preprint arXiv:2507.23208 (2025).arxiv.org/abs/2507.23208
  31. [31]Qijiong Liu, Lu Fan, Zhongzhou Liu, Xiaoyu Dong, Yuankai Luo, Guoyuan An, Nuo Chen, Wei Guo, Yong Liu, and Xiao-Ming Wu. 2026. Accelerating Generative Recommendation via Simple Categorical User Sequence Compression. In Proceedings of the Nineteenth ACM International Conference on Web Search and Data Mining. 1195–1199.
  32. [32]Tianqi Liu, Kairui Fu, Shengyu Zhang, Wenyuan Fan, Zhaocheng Du, Jieming Zhu, Fan Wu, and Fei Wu. 2025. CHORD: Customizing Hybrid-precision On-device Model for Sequential Recommendation with Device-cloud Collaboration. In Proceedings of the 33rd ACM International Conference on Multimedia. Association for Computing Machinery, New York, NY, USA, 6252–6261. doi:10.1145/3746027.3755362
  33. [33]Zheqi Lv, Tianyu Zhan, Wenjie Wang, Xinyu Lin, Shengyu Zhang, Wenqiao Zhang, Jiwei Li, Kun Kuang, and Fei Wu. 2025. Collaboration of Large Language Models and Small Recommendation Models for Device-Cloud Recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1. 962–973.DOI
  34. [34]Zheqi Lv, Wenqiao Zhang, Zhengyu Chen, Shengyu Zhang, and Kun Kuang. 2024. Intelligent model update strategy for sequential recommendation. In Proceedings of the ACM Web Conference 2024. 3117–3128.DOI
  35. [35]Nattaya Mairittha, Tittaya Mairittha, and Sozo Inoue. 2020. Improving activity data collection with on-device personalization using fine-tuning. In UbiComp/ISWC. 255–260.
  36. [36]Xufeng Qian, Yue Xu, Fuyu Lv, Shengyu Zhang, Ziwen Jiang, Qingwen Liu, Xiaoyi Zeng, Tat-Seng Chua, and Fei Wu. 2022. Intelligent request strategy design in recommender system. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3772–3782.arxiv.org/abs/2206.12296
  37. [37]Yifan Shao, Peilin Zhou, Shoujun Wang, Weizhi Zhang, Xu Cai, and Sunghun Kim. 2025. DTRec: Learning Dynamic Reasoning Trajectories for Sequential Recommendation. arXiv preprint arXiv:2512.14036 (2025).arxiv.org/abs/2512.14036
  38. [38]Yehjin Shin, Jeonghwan Choi, Hyowon Wi, and Noseong Park. 2024. An attentive inductive bias for sequential recommendation beyond the self-attention. In Proceedings of the AAAI conference on artificial intelligence, Vol. 38. 8984–8992.
  39. [39]Kartik Sreenivasan, J-yong Sohn, Liu Yang, Matthew Grinde, Alliot Nagle, Hongyi Wang, Eric Xing, Kangwook Lee, and Dimitris Papailiopoulos. 2022. Rare gems: Finding lottery tickets at initialization. Advances in Neural Information Processing Systems 35 (2022), 14529–14540.arxiv.org/abs/2202.12002
  40. [40]Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19). Association for Computing Machinery, New York, NY, USA, 1441–1450. doi:10.1145/3357384.3357895DOI
  41. [41]Qiaoyu Tan, Jianwei Zhang, Jiangchao Yao, Ninghao Liu, Jingren Zhou, Hongxia Yang, and Xia Hu. 2021. Sparse-interest network for sequential recommendation. In Proceedings of the 14th ACM international conference on web search and data mining. 598–606.arxiv.org/abs/2102.09267
  42. [42]Louis Leon Thurstone. 1927. A Law of Comparative Judgment. Psychological Review 34, 4 (1927), 273–286.
  43. [43]Jianling Wang, Raphael Louca, Diane Hu, Caitlin Cellier, James Caverlee, and Lijangjie Hong. 2020. Time to shop for valentine’s day: Shopping occasions and sequential recommendation in e-commerce. In Proceedings of the 13th international conference on web search and data mining. 645–653.DOI
  44. [44]Qinyong Wang, Hongzhi Yin, Tong Chen, Junliang Yu, Alexander Zhou, and Xiangliang Zhang. 2021. Fast-Adapting and Privacy-Preserving Federated Recommender System. The VLDB Journal 31, 5 (2021), 877–896.arxiv.org/abs/2104.00919
  45. [45]Xiaofei Wang, Yiwen Han, Victor CM Leung, Dusit Niyato, Xueqiang Yan, and Xu Chen. 2020. Convergence of edge computing and deep learning: A comprehensive survey. IEEE communications surveys & tutorials 22, 2 (2020), 869–904.DOI
  46. [46]Yunjia Xi, Weiwen Liu, Yang Wang, Ruiming Tang, Weinan Zhang, Yue Zhu, Rui Zhang, and Yong Yu. 2023. On-device integrated re-ranking with heterogeneous behavior modeling. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5225–5236.DOI
  47. [47]Xin Xia, Hongzhi Yin, Junliang Yu, Qinyong Wang, Guangdong Xu, and Quoc Viet Hung Nguyen. 2022. On-device next-item recommendation with self-supervised knowledge distillation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 546–555.arxiv.org/abs/2204.11091
  48. [48]Bencheng Yang, Pengjie Wang, Kai Zhang, Feng Li, Hongbo Deng, Jian Xu, and Bo Zheng. 2022. Apg: Adaptive parameter generation network for click-through rate prediction. Advances in Neural Information Processing Systems 35 (2022), 24740–24752.
  49. [49]Yikai Yan, Chaoyue Niu, Renjie Gu, Fan Wu, Shaojie Tang, Lifeng Hua, Chengfei Lyu, and Guihai Chen. 2022. On-device learning for model personalization with large-scale cloud-coordinated domain adaption. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2180–2190.DOI
  50. [50]Tao Yang, Cuize Han, Chen Luo, Parth Gupta, Jeff M Phillips, and Qingyao Ai. 2024. Mitigating exploitation bias in learning to rank with an uncertainty-aware empirical Bayes approach. In Proceedings of the ACM Web Conference 2024. 1486–1496.arxiv.org/abs/2305.16606
  51. [51]Jiangchao Yao, Feng Wang, Kunyong Jia, Bo Han, Jingren Zhou, and Hongxia Yang. 2021. Device-cloud collaborative learning for recommendation. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 3865–3874.arxiv.org/abs/2104.06624
  52. [52]Deming Ye, Yankai Lin, Yufei Huang, and Maosong Sun. 2021. Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618 (2021).arxiv.org/abs/2105.11618
  53. [53]Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. 2025. ATP-LLAVA: Adaptive Token Pruning for Large Vision Language Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 24972–24982.
  54. [54]Hongzhi Yin, Lianq Qiu, Tong Chen, Wei Yuan, Ruiqi Zheng, Jing Long, Xin Xia, Yuhui Shi, and Chengqi Zhang. 2024. On-device recommender systems: A comprehensive survey. arXiv preprint arXiv:2401.11441 (2024).arxiv.org/abs/2401.11441
  55. [55]Qihang Yu, Kairui Fu, Zheqi Lv, Shengyu Zhang, Xinhui Wu, Chen Lin, Feng Wei, Bo Zheng, and Fei Wu. 2026. ThinkRec: Thinking-based Recommendation via LLM. In Proceedings of the ACM Web Conference 2026. 5698–5709.DOI
  56. [56]Tianyuzhuan Zhu, Gui Ling, Ran Tao, Kunhai Lin, Tong Xiong, Wang Yang, Kaixuan Zhang, Zhihong Chen, Yuliang Yan, Dan Ou, Shengyu Zhang, and Haihong Tang. 2026. SLE: Search-to-List Item Distillation for Single-Rollout Generative Retrieval. arXiv preprint arXiv:2610.34764 (2026).
  57. [57]Tianyuzhuan Shen, Shengyu Zhang, Zheqi Lv, Jieming Zhu, Jiwei Li, Fan Wu, and Fei Wu. 2025. Device-Cloud Collaborative Correction for On-Device Recommendation. In Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence. 3615–3623.arxiv.org/abs/2506.12687
  58. [58]Qiang Zhang, Hanchao Yu, Ivan Ji, Chen Yuan, Yi Zhang, Chihuang Liu, Xiaolong Wang, Christopher E Lambert, Ren Chen, Chen Kovacs, et al. 2025. Efficient Sequential Recommendation for Long Term User Interest Via Personalization. In 2025 IEEE International Conference on Data Mining (ICDM). IEEE, 913–922.
  59. [59]Wei Zhao, Benyou Wang, Min Yang, Jianbo Ye, Zhou Zhaoqin, Xiaojun Chen, and Ying Shen. 2019. Leveraging long and short-term information in content-aware movie recommendation via adversarial training. IEEE transactions on cybernetics 50, 11 (2019), 4680–4693.DOI
  60. [60]Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1059–1068.arxiv.org/abs/1706.06978
  61. [61]Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization. In Proceedings of the 29th ACM international conference on information & knowledge management. 1893–1902.
  62. [62]Feng Zha, Chaochao Chen, Yan Wang, Guan Feng Liu, and Xiaolin Zheng. 2019. DTCDR: A Framework for Dual-Target Cross-Domain Recommendation. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19). Association for Computing Machinery,

Paper details

vlnc:W_jaycv3w6
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents